NVIDIA Unveils Blackwell Architecture and GB200 NVL72: The Trillion-Parameter AI Computing Monster

On March 18, 2024, before an arena of more than 18,000 technology luminaries inside the SAP Center in San Jose, California, NVIDIA CEO Jensen Huang unveiled what is celebrated as the most formidable computing engine in human history: the Blackwell architecture, headlined by the NVIDIA B200 GPU and the liquid-cooled GB200 NVL72 rack-scale supercomputer. Hailed by international press as the “Woodstock of AI,” this milestone accelerated humanity’s computing frontier from hundred-billion parameter models straight into the trillion-parameter multimodal epoch.

Fusing Silicon: 208 Billion Transistors at 10 TB/s

Named in honor of pioneering mathematician David Harold Blackwell, the architecture solved the hard physical boundary of modern photolithographic reticle limits via an audacious dual-die topology:

• Two Monolithic Dies Interconnected as One (208 Billion Transistors): Fabricated on TSMC’s 4NP node, two maximum-reticle dies were bound by a 10 TB/s bidirectional chip-to-chip interconnect (NV-HPI). The internal fabric delivers coherent cache line-rate latency, functioning transparently to developers as a singular, unified super-GPU.
• Second-Generation Transformer Engine and Native FP4 Precision: Pioneering micro-tensor scaling math directly in hardware, the silicon enabled native FP4 (4-bit floating point) arithmetic. This quadrupled computational density without precision loss, yielding an astronomical 20 PFLOPS of FP4 tensor throughput per B200 accelerator.
• 192GB HBM3e Memory at 8.0 TB/s: Eight 24GB HBM3e memory stacks surrounded the dual silicon cores, delivering an astounding 8.0 TB/s of memory bandwidth.

The Liquid-Cooled GB200 NVL72: The Entire Rack is a Single GPU

The technological showstopper of GTC 2024 was the GB200 NVL72, a liquid-cooled rack-scale computational universe:

1. 36 Grace CPUs Coupled with 72 Blackwell GPUs: The massive high-density cabinet united 36 Arm-based Grace processors and 72 B200 accelerators, operating within a high-efficiency closed-loop liquid cooling envelope.
2. Fifth-Generation NVLink Copper Backbone: Employing nearly two miles of internal high-speed copper cabling and nine NVLink Switch processors, all 72 GPUs communicated over a unified 130 TB/s interconnect fabric, yielding 30TB of unified shared memory.
3. Training 1.8-Trillion Parameter Models at 1/25th the Power: Training a 1.8-trillion parameter mixture-of-experts model previously required 8,000 Hopper GPUs consuming 15 Megawatts over 90 days. Blackwell completed the task using merely 2,000 GPUs on only 4 Megawatts of power—a historic 25x reduction in energy consumption.

Historical Perspective: Redefining Hyperscale Infrastructure

The launch of Blackwell established the new planetary baseline for artificial intelligence supercomputing. Tech giants including Microsoft, Amazon, Google, Meta, Oracle, Tesla, and OpenAI committed billions in capital expenditures to deploy Blackwell mega-clusters, cementing NVIDIA as the primary industrial architect of the AGI transition.